Ocean Wave Forecasting with Deep Learning as Alternative to Conventional Models

TL;DR

OCN introduces OceanCastNet, a global deep-learning wave forecasting model that ingests wind forcing and historical wave states to predict $H_s$, $T_m$, and $\theta_m$ with accuracy comparable to ECWAM while delivering orders-of-magnitude faster forecasts. The approach uses an adaptive Fourier neural operator (AFNO) backbone in an auto-regressive, multi-timestep framework, trained on ERA5 and validated against NDBC buoy and Jason-3 satellite data, including extreme weather cases like Typhoon Goni. Across ERA5-based idealized forecasts and real-world comparisons, OCN achieves high anomaly correlation and low RMSE, with robust performance at 360-hour lead times and noticeable regional strengths (West Pacific) relative to ECWAM. The study demonstrates substantial computational efficiency gains and suggests DL-based wave forecasting as a practical, scalable alternative for operational sea-state prediction and real-time ensemble applications.

Abstract

This study presents OceanCastNet (OCN), a machine learning approach for wave forecasting that incorporates wind and wave fields to predict significant wave height, mean wave period, and mean wave direction.We evaluate OCN's performance against the operational ECWAM model using two independent datasets: NDBC buoy and Jason-3 satellite observations. NDBC station validation indicates OCN performs better at 24 stations compared to ECWAM's 10 stations, and Jason-3 satellite validation confirms similar accuracy across 228-hour forecasts. OCN successfully captures wave patterns during extreme weather conditions, demonstrated through Typhoon Goni with prediction errors typically within $\pm$0.5 m. The approach also offers computational efficiency advantages. The results suggest that machine learning approaches can achieve performance comparable to conventional wave forecasting systems for operational wave prediction applications.

Figures (7)

• Figure 1: OceanCastNet Architecture and Auto-Regressive Workflow. (a) Model Architecture for a single forecast step. The input tensor, containing wave and wind data from multiple time steps, is partitioned into non-overlapping patches. These patches are linearly embedded into high-dimensional tokens with added positional encoding. The tokens are then processed by a series of 12 AFNO Layers, where information is mixed globally via a Fourier transform-based spatial mixer and locally via an MLP-based channel mixer. Finally, a linear decoder reconstructs the output grid of predicted wave parameters, which is then passed to a masked loss function for error calculation during training. (b) Auto-Regressive Forecasting Workflow. The model is initialized with a known wave state (t-1, t) and corresponding wind forcing. In "Forecast step 1," the model (using the architecture from panel (a) predicts the wave state at t+1. For "Forecast step 2," this predicted state becomes part of the new input, and the model is provided with new, externally-supplied future wind forcing (t+3) to predict the wave state at t+2. This iterative process continues for the entire forecast duration.
• Figure 2: Global performance analysis of the wave prediction model based on 36 forecast cases throughout the year. (a-b, e-f, i-j) Spatial distribution of mean error (ME) for significant wave height ($H_s$), mean wave period ($T_m$), and mean wave direction ($\theta_m$) at 6-hour and 360-hour lead times. Red indicates positive bias (over-prediction) while blue indicates negative bias (under-prediction). (c-d, g-h, k-l) Temporal evolution of anomaly correlation coefficient (ACC) and root mean square error (RMSE) for each parameter over the 360-hour forecast period. Solid lines represent the mean across all 36 forecast cases, while shaded areas represent the interquartile range (25th to 75th percentiles).
• Figure 3: Model performance during the peak intensity phase of Typhoon Goni (October 30 - November 1, 2020). The figure presents a time series comparison of $H_s$ fields: (a,d,g,j,m,p) model predictions, (b,e,h,k,n,q) ERA5 reference data, and (c,f,i,l,o,r) bias maps (model minus ERA5) at six sequential time points. Bias maps use a diverging colormap where red indicates over-prediction and blue indicates under-prediction.
• Figure 4: Global distribution of 56 National Data Buoy Center (NDBC) stations used for model validation in 2020. Red dots indicate the locations of the selected buoy stations.
• Figure 5: Model performance evaluation against NDBC buoy data for $H_s$ in 2020. (a) Yearly average correlation coefficient and (b) RMSE over 60 time steps across 36 forecast cases for both OCN (red) and ERA5 (blue). Solid lines represent mean values while shaded areas depict the interquartile range (25th to 75th percentiles). (c) Scatter plot of OCN predicted vs. observed $H_s$ and (d) ERA5 vs. observed $H_s$ for all stations and all forecast cases. (e) Time series comparison of observed (black), OCN (red), and ERA5 (blue) $H_s$ for Station 41047 for the forecast case initiated on December 12th, 2020.